mirror_zfs/lib/libzpool/taskq.c
ednadolski-ix 3bd4df3841
Improve ZFS objset sync parallelism
As part of transaction group commit, dsl_pool_sync() sequentially calls
dsl_dataset_sync() for each dirty dataset, which subsequently calls
dmu_objset_sync().  dmu_objset_sync() in turn uses up to 75% of CPU
cores to run sync_dnodes_task() in taskq threads to sync the dirty
dnodes (files).

There are two problems:

1. Each ZVOL in a pool is a separate dataset/objset having a single
   dnode.  This means the objsets are synchronized serially, which
   leads to a bottleneck of ~330K blocks written per second per pool.

2. In the case of multiple dirty dnodes/files on a dataset/objset on a
   big system they will be sync'd in parallel taskq threads. However,
   it is inefficient to to use 75% of CPU cores of a big system to do
   that, because of (a) bottlenecks on a single write issue taskq, and
   (b) allocation throttling.  In addition, if not for the allocation
   throttling sorting write requests by bookmarks (logical address),
   writes for different files may reach space allocators interleaved,
   leading to unwanted fragmentation.

The solution to both problems is to always sync no more and (if
possible) no fewer dnodes at the same time than there are allocators
the pool.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Edmund Nadolski <edmund.nadolski@ixsystems.com>
Closes #15197
2023-11-06 10:38:42 -08:00

417 lines
9.7 KiB
C

/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or https://opensource.org/licenses/CDDL-1.0.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright 2010 Sun Microsystems, Inc. All rights reserved.
* Use is subject to license terms.
*/
/*
* Copyright 2011 Nexenta Systems, Inc. All rights reserved.
* Copyright 2012 Garrett D'Amore <garrett@damore.org>. All rights reserved.
* Copyright (c) 2014 by Delphix. All rights reserved.
*/
#include <sys/zfs_context.h>
int taskq_now;
taskq_t *system_taskq;
taskq_t *system_delay_taskq;
static pthread_key_t taskq_tsd;
#define TASKQ_ACTIVE 0x00010000
static taskq_ent_t *
task_alloc(taskq_t *tq, int tqflags)
{
taskq_ent_t *t;
int rv;
again: if ((t = tq->tq_freelist) != NULL && tq->tq_nalloc >= tq->tq_minalloc) {
ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
tq->tq_freelist = t->tqent_next;
} else {
if (tq->tq_nalloc >= tq->tq_maxalloc) {
if (!(tqflags & KM_SLEEP))
return (NULL);
/*
* We don't want to exceed tq_maxalloc, but we can't
* wait for other tasks to complete (and thus free up
* task structures) without risking deadlock with
* the caller. So, we just delay for one second
* to throttle the allocation rate. If we have tasks
* complete before one second timeout expires then
* taskq_ent_free will signal us and we will
* immediately retry the allocation.
*/
tq->tq_maxalloc_wait++;
rv = cv_timedwait(&tq->tq_maxalloc_cv,
&tq->tq_lock, ddi_get_lbolt() + hz);
tq->tq_maxalloc_wait--;
if (rv > 0)
goto again; /* signaled */
}
mutex_exit(&tq->tq_lock);
t = kmem_alloc(sizeof (taskq_ent_t), tqflags);
mutex_enter(&tq->tq_lock);
if (t != NULL) {
/* Make sure we start without any flags */
t->tqent_flags = 0;
tq->tq_nalloc++;
}
}
return (t);
}
static void
task_free(taskq_t *tq, taskq_ent_t *t)
{
if (tq->tq_nalloc <= tq->tq_minalloc) {
t->tqent_next = tq->tq_freelist;
tq->tq_freelist = t;
} else {
tq->tq_nalloc--;
mutex_exit(&tq->tq_lock);
kmem_free(t, sizeof (taskq_ent_t));
mutex_enter(&tq->tq_lock);
}
if (tq->tq_maxalloc_wait)
cv_signal(&tq->tq_maxalloc_cv);
}
taskqid_t
taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t tqflags)
{
taskq_ent_t *t;
if (taskq_now) {
func(arg);
return (1);
}
mutex_enter(&tq->tq_lock);
ASSERT(tq->tq_flags & TASKQ_ACTIVE);
if ((t = task_alloc(tq, tqflags)) == NULL) {
mutex_exit(&tq->tq_lock);
return (0);
}
if (tqflags & TQ_FRONT) {
t->tqent_next = tq->tq_task.tqent_next;
t->tqent_prev = &tq->tq_task;
} else {
t->tqent_next = &tq->tq_task;
t->tqent_prev = tq->tq_task.tqent_prev;
}
t->tqent_next->tqent_prev = t;
t->tqent_prev->tqent_next = t;
t->tqent_func = func;
t->tqent_arg = arg;
t->tqent_flags = 0;
cv_signal(&tq->tq_dispatch_cv);
mutex_exit(&tq->tq_lock);
return (1);
}
taskqid_t
taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg, uint_t tqflags,
clock_t expire_time)
{
(void) tq, (void) func, (void) arg, (void) tqflags, (void) expire_time;
return (0);
}
int
taskq_empty_ent(taskq_ent_t *t)
{
return (t->tqent_next == NULL);
}
void
taskq_init_ent(taskq_ent_t *t)
{
t->tqent_next = NULL;
t->tqent_prev = NULL;
t->tqent_func = NULL;
t->tqent_arg = NULL;
t->tqent_flags = 0;
}
void
taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
taskq_ent_t *t)
{
ASSERT(func != NULL);
/*
* Mark it as a prealloc'd task. This is important
* to ensure that we don't free it later.
*/
t->tqent_flags |= TQENT_FLAG_PREALLOC;
/*
* Enqueue the task to the underlying queue.
*/
mutex_enter(&tq->tq_lock);
if (flags & TQ_FRONT) {
t->tqent_next = tq->tq_task.tqent_next;
t->tqent_prev = &tq->tq_task;
} else {
t->tqent_next = &tq->tq_task;
t->tqent_prev = tq->tq_task.tqent_prev;
}
t->tqent_next->tqent_prev = t;
t->tqent_prev->tqent_next = t;
t->tqent_func = func;
t->tqent_arg = arg;
cv_signal(&tq->tq_dispatch_cv);
mutex_exit(&tq->tq_lock);
}
void
taskq_wait(taskq_t *tq)
{
mutex_enter(&tq->tq_lock);
while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
mutex_exit(&tq->tq_lock);
}
void
taskq_wait_id(taskq_t *tq, taskqid_t id)
{
(void) id;
taskq_wait(tq);
}
void
taskq_wait_outstanding(taskq_t *tq, taskqid_t id)
{
(void) id;
taskq_wait(tq);
}
static __attribute__((noreturn)) void
taskq_thread(void *arg)
{
taskq_t *tq = arg;
taskq_ent_t *t;
boolean_t prealloc;
VERIFY0(pthread_setspecific(taskq_tsd, tq));
mutex_enter(&tq->tq_lock);
while (tq->tq_flags & TASKQ_ACTIVE) {
if ((t = tq->tq_task.tqent_next) == &tq->tq_task) {
if (--tq->tq_active == 0)
cv_broadcast(&tq->tq_wait_cv);
cv_wait(&tq->tq_dispatch_cv, &tq->tq_lock);
tq->tq_active++;
continue;
}
t->tqent_prev->tqent_next = t->tqent_next;
t->tqent_next->tqent_prev = t->tqent_prev;
t->tqent_next = NULL;
t->tqent_prev = NULL;
prealloc = t->tqent_flags & TQENT_FLAG_PREALLOC;
mutex_exit(&tq->tq_lock);
rw_enter(&tq->tq_threadlock, RW_READER);
t->tqent_func(t->tqent_arg);
rw_exit(&tq->tq_threadlock);
mutex_enter(&tq->tq_lock);
if (!prealloc)
task_free(tq, t);
}
tq->tq_nthreads--;
cv_broadcast(&tq->tq_wait_cv);
mutex_exit(&tq->tq_lock);
thread_exit();
}
taskq_t *
taskq_create(const char *name, int nthreads, pri_t pri,
int minalloc, int maxalloc, uint_t flags)
{
(void) pri;
taskq_t *tq = kmem_zalloc(sizeof (taskq_t), KM_SLEEP);
int t;
if (flags & TASKQ_THREADS_CPU_PCT) {
int pct;
ASSERT3S(nthreads, >=, 0);
ASSERT3S(nthreads, <=, 100);
pct = MIN(nthreads, 100);
pct = MAX(pct, 0);
nthreads = (sysconf(_SC_NPROCESSORS_ONLN) * pct) / 100;
nthreads = MAX(nthreads, 1); /* need at least 1 thread */
} else {
ASSERT3S(nthreads, >=, 1);
}
rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
cv_init(&tq->tq_maxalloc_cv, NULL, CV_DEFAULT, NULL);
(void) strlcpy(tq->tq_name, name, sizeof (tq->tq_name));
tq->tq_flags = flags | TASKQ_ACTIVE;
tq->tq_active = nthreads;
tq->tq_nthreads = nthreads;
tq->tq_minalloc = minalloc;
tq->tq_maxalloc = maxalloc;
tq->tq_task.tqent_next = &tq->tq_task;
tq->tq_task.tqent_prev = &tq->tq_task;
tq->tq_threadlist = kmem_alloc(nthreads * sizeof (kthread_t *),
KM_SLEEP);
if (flags & TASKQ_PREPOPULATE) {
mutex_enter(&tq->tq_lock);
while (minalloc-- > 0)
task_free(tq, task_alloc(tq, KM_SLEEP));
mutex_exit(&tq->tq_lock);
}
for (t = 0; t < nthreads; t++)
VERIFY((tq->tq_threadlist[t] = thread_create(NULL, 0,
taskq_thread, tq, 0, &p0, TS_RUN, pri)) != NULL);
return (tq);
}
void
taskq_destroy(taskq_t *tq)
{
int nthreads = tq->tq_nthreads;
taskq_wait(tq);
mutex_enter(&tq->tq_lock);
tq->tq_flags &= ~TASKQ_ACTIVE;
cv_broadcast(&tq->tq_dispatch_cv);
while (tq->tq_nthreads != 0)
cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
tq->tq_minalloc = 0;
while (tq->tq_nalloc != 0) {
ASSERT(tq->tq_freelist != NULL);
taskq_ent_t *tqent_nexttq = tq->tq_freelist->tqent_next;
task_free(tq, tq->tq_freelist);
tq->tq_freelist = tqent_nexttq;
}
mutex_exit(&tq->tq_lock);
kmem_free(tq->tq_threadlist, nthreads * sizeof (kthread_t *));
rw_destroy(&tq->tq_threadlock);
mutex_destroy(&tq->tq_lock);
cv_destroy(&tq->tq_dispatch_cv);
cv_destroy(&tq->tq_wait_cv);
cv_destroy(&tq->tq_maxalloc_cv);
kmem_free(tq, sizeof (taskq_t));
}
/*
* Create a taskq with a specified number of pool threads. Allocate
* and return an array of nthreads kthread_t pointers, one for each
* thread in the pool. The array is not ordered and must be freed
* by the caller.
*/
taskq_t *
taskq_create_synced(const char *name, int nthreads, pri_t pri,
int minalloc, int maxalloc, uint_t flags, kthread_t ***ktpp)
{
taskq_t *tq;
kthread_t **kthreads = kmem_zalloc(sizeof (*kthreads) * nthreads,
KM_SLEEP);
(void) pri; (void) minalloc; (void) maxalloc;
flags &= ~(TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT | TASKQ_DC_BATCH);
tq = taskq_create(name, nthreads, minclsyspri, nthreads, INT_MAX,
flags | TASKQ_PREPOPULATE);
VERIFY(tq != NULL);
VERIFY(tq->tq_nthreads == nthreads);
for (int i = 0; i < nthreads; i++) {
kthreads[i] = tq->tq_threadlist[i];
}
*ktpp = kthreads;
return (tq);
}
int
taskq_member(taskq_t *tq, kthread_t *t)
{
int i;
if (taskq_now)
return (1);
for (i = 0; i < tq->tq_nthreads; i++)
if (tq->tq_threadlist[i] == t)
return (1);
return (0);
}
taskq_t *
taskq_of_curthread(void)
{
return (pthread_getspecific(taskq_tsd));
}
int
taskq_cancel_id(taskq_t *tq, taskqid_t id)
{
(void) tq, (void) id;
return (ENOENT);
}
void
system_taskq_init(void)
{
VERIFY0(pthread_key_create(&taskq_tsd, NULL));
system_taskq = taskq_create("system_taskq", 64, maxclsyspri, 4, 512,
TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
system_delay_taskq = taskq_create("delay_taskq", 4, maxclsyspri, 4,
512, TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
}
void
system_taskq_fini(void)
{
taskq_destroy(system_taskq);
system_taskq = NULL; /* defensive */
taskq_destroy(system_delay_taskq);
system_delay_taskq = NULL;
VERIFY0(pthread_key_delete(taskq_tsd));
}